G7 Leaders Tell Trump They Want U.S. AI but Fear the Off Switch. Amodei Pitches a Coalition.

Since the Trump administration issued an export-control directive that forced Anthropic to block access to Fable 5 and Mythos 5 for all users, including its own employees and U.S.-based foreign nationals, the diplomatic fallout has moved faster than any legal clarification.

On Wednesday, June 17, Anthropic CEO Dario Amodei sat down at a closed-door lunch in Évian-les-Bains, France, alongside approximately a dozen other tech executives, including OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis, plus the heads of state of every G7 country and President Donald Trump, according to CNBC.

Amodei and Hassabis both called for a U.S.-led coalition to set international rules and standards for AI, according to two people with knowledge of the meeting cited by CNBC. Canadian Prime Minister Mark Carney agreed the U.S. should lead such a coalition, per CNBC's sources.

Amodei specifically proposed that the coalition cover structured access to frontier models, chip trade that excludes China, and international cooperation on AI risks in cyber, bioterrorism, and intelligence, CNBC reported.

French President Emmanuel Macron used the same lunch to deliver a pointed warning: if the U.S. can "from one day to the next turn off the switch," it damages not just allied economies but the AI companies themselves, according to TechCrunch. Indian PM Narendra Modi echoed the concern, telling the group that democratic nations must have reliable access to top AI models to protect critical infrastructure, per the Financial Times as cited by TechCrunch.

The allies aren't wrong on the economics. Any government or company that has built workflows on U.S. AI infrastructure now has concrete evidence that access can vanish overnight, for reasons they may not be told. Aidan Gomez, CEO of Canadian enterprise AI firm Cohere, put it plainly in a statement shared with TechCrunch: "Digital sovereignty is not just about market competition or any one company or nation. It's about who controls the foundational technology that will shape our economic security and national sovereignty for decades to come."

The G7 meeting also surfaced a partial answer: a proposed "trusted partners" scheme that would give non-U.S. nations and companies structured access to frontier models from firms like Anthropic and OpenAI, bypassing blanket restrictions, as long as the access was used to build defenses against rivals like China. TechCrunch noted the scheme's limits are unresolved. It is not clear whether a startup in Paris or Bangalore whose product was disrupted last Friday would qualify.

The Legal Void at the Center of This

The strongest concern about the administration's conduct here isn't political. It's structural. Export controls have historically applied to discrete goods that move across borders: hardware, software files, source code. Anthropic's models weren't being shipped anywhere. They were being accessed remotely through a cloud service.

The Verge reported Wednesday that legal experts consider the episode unprecedented. "To my knowledge, this is the first time U.S. export controls have been used to control access to an AI model in this way," one expert told The Verge. The administration has NOT publicly identified the specific legal authority it used. Anthropic's own statement says the government cited "national security authorities" and characterized it as "an export control directive," but that's Anthropic's characterization, not a White House legal memo.

Biden-era rules had begun covering AI model weights, the core data that makes a model function and can be copied elsewhere. The Trump administration scrapped that framework early in the second term. What replaced it, legally, remains unclear.

The Case for Acting

Cybersecurity concerns about frontier models with advanced capabilities are not invented. Amazon flagged to the White House that certain safety guardrails in Fable 5 and Mythos 5 could be bypassed, and that groups linked to China had potentially exploited this, according to TechCrunch. If a model can be jailbroken in ways that matter for national security, waiting for a perfect legal framework before acting carries its own risk.

Anthropic itself disputed the severity, stating that the flagged jailbreak did not allow users to circumvent all of its safeguards, according to The Verge. Cybersecurity experts cited by TechCrunch argued that the capabilities cited by the government are also present in models that remain freely available, including from OpenAI. That asymmetry—Anthropic's models frozen while competitors stay online—is a legitimate question the administration has not addressed publicly.

The Factionalism Variable

The Verge's Tina Nguyen reported Wednesday that the omnishambles around the original Friday-night directive reflects competing factions inside the Trump administration rather than a single coordinated strategy, a pattern she tracked extensively during Trump's first term. Which faction drove the Anthropic order, and whether Treasury, Commerce, or State were fully aligned, has NOT been confirmed by any named source. Scott Bessent, Howard Lutnik, and Marco Rubio were all present at the G7 lunch Wednesday, per CNBC, which may indicate some internal alignment is being worked out in real time.

What Comes Next

The "trusted partners" framework discussed at the G7 is still conceptual. No timeline, membership criteria, or governance structure has been announced. Fable 5 and Mythos 5 remain offline as of June 17. The administration has not indicated when or whether it will provide a public legal rationale for the directive. Whether Congress moves to codify export-control authority over AI service access, or whether the courts get involved if Anthropic chooses to challenge the order, are the two concrete next questions this story is heading toward.